Ignore Prior Instructions: AI Still Befuddled by Basic Reasoning

Ignore Prior Instructions: AI Still Befuddled by Basic Reasoning

A fact is a fact is a fact. But for a large language model (LLM), a fact is what someone says is a fact, if, in fact, they say it sternly enough. The CTO of Microsoft Azure had select words to share about the state of security within AI operations, which obviously needs work. And while he covered the latest state of the art in prompt injections, jailbreaks and hallucinations, he also discussed a set of concerns around the LLMs basic ability to r…